The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection attacks by ingesting data from numerous external scientific sources.
Ingestion points: Data is fetched from sources such as PubMed (via PubMed_search), NCBI GEO (via GEO_download_dataset), UniProt (via UniProt_get_sequence), and OpenTargets (via OpenTargets_get_associated_targets_by_disease_efoId) as detailed in references/tool-composition.md and references/tool-execution.md.
Boundary markers: The skill documentation and provided example scripts do not specify the use of delimiters or specific instructions to prevent the agent from executing instructions that might be embedded within the retrieved scientific data.
Capability inventory: The skill provides significant capabilities, including executing a wide variety of tools through tu.run(), writing results to the local filesystem using the save_to_file hook (documented in references/api_reference.md), and performing LLM-based summarization of external content.
Sanitization: There is no documentation suggesting the use of sanitization or validation of the content retrieved from external databases before it is passed to the LLM or written to files, which could lead to malicious instructions being processed.

tooluniverse